Numerous embodiments are disclosed for a high voltage generation algorithm and system for generating high voltages necessary for a particular programming operation in analog neural memory used in a deep learning artificial neural network. Compensation measures can be utilized that compensate for changes in voltage or current as the number of cells being programmed changes.
Legal claims defining the scope of protection. Each claim is shown in both the original legal language and a plain English translation.
2. The method of claim 1, wherein the array is a vector matrix multiplier.
A vector matrix multiplier is a specialized hardware accelerator designed to efficiently perform matrix-vector multiplication operations, which are fundamental in machine learning, signal processing, and other computational tasks. Traditional processors often struggle with the high computational demands of these operations, leading to bottlenecks in performance and energy efficiency. The vector matrix multiplier addresses this problem by providing a dedicated architecture optimized for matrix-vector multiplication, significantly improving throughput and reducing latency compared to general-purpose processors. The vector matrix multiplier includes an array of processing elements arranged in a matrix configuration, where each processing element is capable of performing multiply-accumulate operations. Input vectors are distributed across the array, and the matrix elements are stored in a memory structure that allows for parallel access. The multiplier supports configurable precision, enabling it to handle both low-precision and high-precision arithmetic as required by different applications. Additionally, the architecture may include mechanisms for data reordering and pipelining to further enhance efficiency. The multiplier can be integrated into larger systems, such as neural network accelerators or digital signal processors, to offload matrix-vector multiplication tasks. By leveraging parallelism and specialized hardware, it achieves higher performance per watt, making it suitable for energy-constrained environments like edge devices and mobile applications. The design may also incorporate error correction and fault tolerance features to ensure reliability in demanding computational workloads.
3. The method of claim 2, where the selected memory cells are split-gate flash memory cells.
This invention relates to memory systems, specifically flash memory, addressing challenges in data storage and retrieval efficiency. The method involves selecting memory cells for programming or reading operations, with a focus on optimizing performance and reliability. The selected memory cells are split-gate flash memory cells, a type of non-volatile memory known for improved data retention and reduced interference between adjacent cells. Split-gate flash memory cells use a split-gate structure to enhance control over charge storage and retrieval, minimizing leakage and improving endurance. The method likely involves techniques for managing these cells, such as precise voltage application during programming or reading to ensure accurate data handling. By leveraging split-gate architecture, the invention aims to enhance memory system performance, particularly in applications requiring high reliability and long-term data integrity. The approach may include steps for selecting specific cells, applying appropriate voltages, and verifying data integrity, all tailored to the unique characteristics of split-gate flash memory. This method is particularly useful in embedded systems, solid-state drives, and other storage solutions where durability and efficiency are critical.
4. The method of claim 2, wherein the selected memory cells are stacked-gate flash memory cells.
This invention describes a method for programming analog neural memory cells, relevant for deep learning artificial neural networks. Although the full details of the foundational method (Claim 1) are not provided, this claim specifies that the method involves an array configured as a vector matrix multiplier. For this particular method, the individual memory cells selected and utilized within this vector matrix multiplier are specifically stacked-gate flash memory cells. This means the system leverages stacked-gate flash technology as the non-volatile storage component for its analog neural memory cells within a vector-matrix computation architecture. ERROR (embedding): Error: Failed to save embedding: Could not find the 'embedding' column of 'patent_claims' in the schema cache
6. The method of claim 5, wherein the array is a vector matrix multiplier.
A vector matrix multiplier is a specialized hardware accelerator designed to efficiently perform matrix-vector multiplication operations, which are fundamental in machine learning, signal processing, and other computational tasks. Traditional processors often struggle with the high computational demands of these operations, leading to bottlenecks in performance and energy efficiency. The vector matrix multiplier addresses this by providing a dedicated architecture optimized for matrix-vector multiplication, reducing latency and improving throughput. The multiplier includes an array of processing elements arranged to handle parallel computations, where each element performs a multiply-accumulate operation. Input vectors and matrices are loaded into the array, and the multiplier processes them in parallel, significantly accelerating the computation compared to general-purpose processors. The architecture may include configurable parameters to adapt to different matrix sizes and data types, ensuring flexibility for various applications. Additionally, the multiplier may integrate with memory systems to minimize data movement overhead, further enhancing efficiency. By offloading matrix-vector multiplication tasks from the central processor, the vector matrix multiplier improves overall system performance, reduces power consumption, and enables real-time processing in applications such as neural networks, digital signal processing, and scientific computing. The design may also incorporate error correction mechanisms to ensure computational accuracy, making it suitable for high-reliability environments.
7. The method of claim 6, where the non-volatile memory cells are split-gate flash memory cells.
This invention relates to memory systems, specifically improving the reliability and performance of non-volatile memory cells. The problem addressed is the degradation of memory cells over time, particularly in flash memory, which can lead to data corruption and reduced lifespan. The solution involves using split-gate flash memory cells, which are designed to mitigate charge leakage and improve endurance compared to traditional single-gate designs. Split-gate cells have two gates: a control gate and a select gate, allowing for better control over charge storage and reducing interference between adjacent cells. This design helps maintain data integrity over repeated read/write cycles. The method also includes techniques for managing the memory cells, such as adjusting programming voltages or using error correction, to further enhance reliability. By employing split-gate flash memory cells, the system achieves longer lifespan, higher data retention, and improved performance in non-volatile memory applications. The invention is particularly useful in storage devices where durability and reliability are critical, such as solid-state drives and embedded memory systems.
8. The method of claim 6, wherein the non-volatile memory cells are stacked-gate flash memory cells.
This invention relates to memory systems, specifically methods for operating stacked-gate flash memory cells. Stacked-gate flash memory cells are a type of non-volatile memory where data is stored by trapping charge in a floating gate structure. A common challenge in such systems is efficiently managing read and write operations while maintaining data integrity and performance. The method involves programming and reading data in stacked-gate flash memory cells, where the cells are organized in a memory array. The process includes applying specific voltage levels to control gates and bit lines to selectively program or read individual cells. During programming, a high voltage is applied to the control gate while a lower voltage is applied to the bit line, causing charge to be trapped in the floating gate. For reading, a controlled voltage is applied to the control gate, and the resulting current flow through the cell is measured to determine the stored data state. The method ensures reliable data storage by carefully controlling the voltage levels and timing of operations, preventing unintended charge leakage or disturbance to adjacent cells. It also optimizes performance by minimizing the time required for programming and reading while maintaining accuracy. The technique is particularly useful in high-density memory applications where efficient data handling is critical.
Cooperative Patent Classification codes for this invention. Click any code to explore related patents in that topic.
March 3, 2021
December 6, 2022
Browse 5M+ US patents with plain-English claim translations and AI-generated analysis.